Contractors training OpenAI models were fired for substituting AI for human judgment

Multiple contractors working on OpenAI projects have been fired or removed for using AI to complete tasks that were supposed to provide human judgment, 404 Media found through internal documents and interviews with three contractors. Some projects involve more than 10,000 workers. Two sources worked for Mercor, which says its contracts prohibit LLM use and confirmed that it removes workers when misuse is established.

The reviewers read and rate model outputs, including real ChatGPT conversations in the previously reported Project Lily. Their work is intended to correct properties such as sycophancy and excessive anthropomorphism. Feeding model-generated judgments back into that process can conceal errors, homogenize preferences and make evaluation circular.

Internal guidance reveals how uncertain detection remains. Reviewers are told not to use automated AI detectors because they are unreliable. Instead, they look for repeated wording, characteristic punctuation and implausibly fast completion—signals that can also implicate an efficient or non-native writer. Reviewers are also instructed not to tell workers which clues triggered suspicion.

The incentive problem is straightforward: contractors doing repetitive, low-autonomy work can increase output with the same tools their employer is promoting everywhere else. A reliable training pipeline therefore cannot depend solely on prohibitions. It needs smaller auditable tasks, provenance controls, calibrated spot checks and compensation that does not reward speed at the expense of independent judgment.